Freeway Lane-Level Traffic Flow Prediction Method Based on Multi-Scale Spatial Feature Fusion

Existing studies on freeway traffic flow prediction mainly focus on single cross-sections, without fully considering spatio-temporal correlations across lanes and along upstream downstream segments. The intrinsic relationship between flow and speed is also often neglected. This study proposes a lane-level freeway traffic flow prediction method based on multi-scale spatial feature fusion. The method quantifies and compensates for spatial time-lag effects between adjacent sections, reducing temporal misalignment of upstream and downstream flow sequences. For spatial feature extraction, flow and speed data with time-lag effects removed are integrated and processed through three-scale dual-channel 3D convolutional modules with attention mechanisms. These modules dynamically capture local inter-lane interactions, global propagation patterns between sections, and intrinsic flow-speed dependencies. For temporal modeling, a long short-term memory network is employed to extract global temporal dependencies among multi-scale spatial features, and a fully connected layer generates final predictions. Empirical validation using real PeMS freeway data demonstrates that, in one-step prediction tasks, the proposed method reduces the mean absolute error, root mean square error, and mean absolute percentage error by at least 6.61%, 5.50%, and 8.46% on average compared with the other models. In multi-step prediction, average errors across horizons decrease by up to 14.09%, 15.25%, and 29.16%, confirming the method's effectiveness in capturing fine-grained multi-scale spatio-temporal features and its significant advantage in prediction accuracy. Furthermore, ablation experiments verify that the attention mechanism and the collaborative integration of multi-scale spatio information play crucial roles in improving freeway traffic flow prediction performance.